An affect lexicon is a compendium of lexical entries for affect words with their corresponding parts of speech, af­fect categories, centralities, and intensities. An affect word is any word having an affect-related meaning or connota­tion. Any given affect word may have multiple entries in an affect lexicon, differing by its part of speech and/or its category (Subasic and Huettner, 2001). In these lexicons entries are labeled with their prior polarity (Wilson et al., 2009), i.e. their polarity without any given context or dis­course. An affect lexicon may then be used to compute the frequency of sentiment bearing words in a text. Thereby the attempt usually is to percieve sentences comprising the unambiguous sense of the sentiment bearing word and/or taking into account contextual valence shifters (Polanyi and Zaenen, 2006).

Entries in the dictionary schematically look like shown in

Table 1.

The part of speech tags (POS tags) are given in the form of (Thielen et al., 1999)'s Stuttgart-Tübingen-Tagset (STTS). As POS tags are only provided for the baseforms, they are limited to those of adjectives, adverbs, normal nouns and infinite verbs. The inflections were, where available, re­trieved from an internal database and are not guaranteed to be complete and error-free. Table 2 provides a comprehen­sive overview of the dictionary's content.

SentiWS exploits several sources providing words plus their semantic orientation in one way or the other. All sources required a manual revision as described below.

The weights mentioned above were retrieved utilising a method first suggested by (Church and Hanks, 1990): the so called Pointwise Mutual Information (PMI). This approach was successfully re-used for work related to sentiment anal­ysis - the determination of the semantic orientation and its strength of adjectives - by (Turney, 2002) and (Turney and Littman, 2003). Their general strategy is to infer semantic orientation from semantic association. The semantic orien­tation SO of a given word w is calculated from the strength of its association A with a manually-selected set of positive seed words P minus the strength of its association with a set of negative seed words N (cf. Equation 1).

The word w is classified as having a positive semantic ori­entation when SO-A(w) is positive and a negative semantic orientation when SO-A(w) is negative. The absolute value of SO-A(w) can be considered the strength of its semantic orientation.

Parallel to (Turney and Littman, 2003)'s paradigms we used the following German seed sets Pde and Nde:

gut, schein, richtig, glücklich, erstklassig, positiv, großartig, ausgezeichnet, lieb, exzellent, phantastisch schlecht, unschön, falsch, unglücklich, zweitklassig, negativ, scheiße, minderwertig, böse, armselig, mies

The semantic associations A(w,p) and A(w,n) are then calculated using the PMI. The PMI between two words wi and w2 according to (Church and Hanks, 1990) is defined as given in Equation 4, where P(w) is the probability that w occurs and P (w1&w2) is the probability that w1 and w2 co-occur. These probabilities were estimated using frequencies and co-occurrence statistics on a internal German-language cor­pus consisting of approximately 100 Million sentences. If SO-A(w) of a word that was entered as being positive is negative or vice versa, the word in question is either re­moved, put in the opposite class, or, if after manually revi­sion we find its classification is correct, its weight is set to the minimum weight of its class. All weights are scaled to the interval of [-1; 1] and rounded to 4 decimal places with +1.0 being absolutely positive and -1.0 being absolutely negative.

Very positive words are for example Freude (i.e. joy) with a weight of 0.6502 and perfekt (i.e. perfect) with a weight of 0.7299. Very negative words are for example betrügen (i.e. to betray) with a weight of -0.743 and schädlich (i.e. harmful) with a weight of -0.9269. The distribution of the absolute weights in SentiWS (cf. Figure 1) follows a Zipf-like distribution (Zipf, 1972): Very little word forms have high weights, some word forms have medium weights and a large amount of word forms have little or very little weights.

Just like there was no German-language dictionary for sen­timent analysis, there is no corresponding data set for evalu­ation purposes. In order to evaluate SentiWS' performance we compiled such a data set. We first randomly selected 2,000 sentences from a corpus containing posts from a va­riety of internet fora and then manually categorized them as being positive, negative or neutral. We then randomly selected 160 sentences from each category, resulting in a data set of 480 sentences. The minimum sentence length is 4 word forms, the maximum sentence length is 40 word forms and the approximate average sentence length is 15.97 word forms.

Two human raters, one of which is an author of this pa­per, were then instructed to annotate each sentence regard­ing the prior polarities of each adjective, adverb, noun or verb in it, i.e. they had to decide whether a word form was positive, negative or had no prior polarity. Measuring the raters' overall agreement using Cohens k (Cohen, 1960) in a free-marginal variant (Brennan and Prediger, 1981) the interrater reliability is k free = 0.76 and thus is considered being reliable.

Hereupon the sentences were preprocessed incorporating the Stanford POS Tagger 2.0 and, taking into account the POS tags, the raters' annotations were compared with the entries in SentiWS. Errors induced by the Stanford POS Tagger were excluded and precision P, recall R and f-measure F were calculated as shown in Table 3.

SentiWS' entries were checked against the annotations of rater 1, rater 2 and their consensus, i.e. the annotations in which both raters agreed. All results are given for pos­itive word forms only, negative word forms only and all word forms. Generally SentiWS performs better identify­ing negative word forms (F = 0.86) than it does for posi­tive word forms (F = 0.82), but the overall performance is very promising (P = 0.96, R = 0.74, F = 0.84).

Typical errors that lower the recall are missing words and missing word forms, for example domain-specific terms (e.g. Kantigkeit, i.e. "edgy-ness"), foreign words, collo­quial language (e.g. Ka***, i.e. sh**) and mistypes ormis-spellings. Typical errors that lower the precision include words that are ambiguous and tend to be polar in one sense but not the other.

SentiWS is work in progress and hence far from being fully-fledged and error-free. It will be continuously refined by adding missing words and word forms and removing am­biguous ones. It is furthermore likely that it will be ex­tended by introducing a new dimension indicating subjec­tivity, just like (Esuli and Sebastiani, 2006)'s SentiWordNet does. Apart from that the authors recently took interest in representing more fine-grained emotions, aside from pure polarity (Whitelaw et al., 2005).

We also believe it is necessary to delve into weighting schemes. Although we used the PMI without questioning it, we are very aware of the fact that the weighting itself needs to be evaluated and possibly contrasted with other weight­ing methods (Landauer and Dumais, 1997; Richardson et al., 1994; Budanitsky and Hirst, 2001; Biemann, 2006)

We have in detail presented a German-language affect dic­tionary which attributes each word with its syntactic cate­gory, its inflectional forms, its polarity and its strength. We conducted an evaluation and proved SentiWS being a use­ful resource for sentiment analysis related tasks to build on. As far as we know SentiWS is the first German-language dictionary dedicated to sentiment analysis, opinion mining etc. publicly available and we encourage researchers to use it in composition with the other corpora, tools and webser­vices provided by the Wortschatz project (Quasthoff et al., 2006; Biemann et al., 2007; Biemann et al., 2008; Buchler